AI Trading Book Reference
Home / Regulation / Definitions
Regulation Definitions UK · AU · NZ

What regulators actually mean by "algorithm" in trading

No financial regulator in the UK, Australia or New Zealand defines "AI trading". They define algorithmic trading, and they define it by function. The test is whether a computer determines the parameters of an order with limited or no human intervention — not whether the system uses machine learning, a neural network or a hand-written rule. That single design choice is why AI needed no new rulebook, and it is also why the word "algorithm" in a regulatory document means something narrower than the word "bot" in a forum post.

Published 27 August 2026 · Updated 28 August 2026 · AI Trading Book Editorial · Reading time about 11 minutes

TL;DR
  • The test is order parameters plus limited human intervention. Everything follows from that.
  • No definition names AI — deliberately, because naming a technology dates the rule.
  • Australia's proposed wording mirrors the UK's and would close a gap between two rulebooks.
  • New Zealand has no equivalent defined term that we could locate.
  • Being an algorithm is not the same as being regulated. The definitions bite on firms, not individuals.

Why the definition decides everything else

Every obligation in the algorithmic trading regime — conformance testing, annual self-assessment, stress testing, kill functionality — hangs on whether an activity falls inside one definition. Get the definition wrong and a firm either builds controls it does not need or fails to build ones it does.

For readers who are not firms, the definition matters for a different reason: it explains why claims that a product is "regulated AI" or "FCA-approved for AI trading" are meaningless. There is no AI category to be approved for. What exists is a functional test that would catch a spreadsheet macro placing orders and would not catch a neural network that only produces a chart.

The UK definition

The UK carries the MiFID-derived definition into the FCA Handbook at MAR 7A, with the technical requirements in onshored RTS 6. Algorithmic trading means trading in financial instruments where a computer algorithm automatically determines individual parameters of orders — whether to initiate the order, its timing, price, quantity, or how to manage it after submission — with limited or no human intervention.

Three elements do the work, and each excludes something.

  • "Determines individual parameters of orders." The system must decide the substance of the order. A system that only transmits an order a human specified is outside it.
  • "Automatically." The decision happens without a person making it in that instance.
  • "With limited or no human intervention." A human approving each order before submission generally takes the activity outside the definition, even if a model produced the suggestion.

Systems used only for routing orders to venues, for confirming orders, or for post-trade processing are generally outside the definition, because none of them determines order parameters.

The functional test, as drafted. Note that meeting it does not by itself make an individual regulated. Does a computer determine parameters of the order? no Outside the definition Routing, confirmation, post-trade yes Limited or no human intervention in that decision? no Outside the definition Human approves each order yes Algorithmic trading Whether or not any AI is involved If you are an authorised firm MAR 7A and RTS 6 obligations apply Note the absent question: "Is it AI?" appears nowhere in the test. Neither does "machine learning" or "neural network".
Compiled from the MiFID-derived definition carried into MAR 7A. Simplified for readability; firms should work from the rule text.

High-frequency trading is a narrower box inside it

The same framework defines a high-frequency algorithmic trading technique as a subset of algorithmic trading, identified by three characteristics operating together: infrastructure intended to minimise network and other latencies, including co-location, proximity hosting or high-speed direct electronic access; system determination of order initiation, generation, routing or execution without human intervention for individual orders; and high message intraday rates constituting orders, quotes or cancellations.

The practical effect is a nested structure. All high-frequency trading is algorithmic trading; most algorithmic trading is not high-frequency. A retail bot placing a few orders a day sits in the outer box and nowhere near the inner one — which matters because much of the public anxiety about "algorithms" is really about the inner box.

Australia's proposed definition

Australia currently has no equivalent defined term. CP 386, published 27 August 2025, proposes one: a Trading Algorithm, described in the draft as a computer algorithm that automatically determines, with limited or no human intervention, one or more parameters of an order. The wording tracks the UK formulation closely, which is consistent with ASIC's stated objective of aligning with IOSCO principles and international practice.

CP 386 also proposes a second and wider concept. The existing rules distinguish between designated trading representatives and automated order processing — a split reflecting an era when a human keying an order and a machine generating one were meaningfully different activities. The proposals would replace both with a single notion of Trading Systems, described as any system for submitting trading messages into a trading platform.

The two terms do different jobs. Trading Systems is broad and captures the pathway an order takes. Trading Algorithm is narrow and captures the decision-making, attracting the additional controls including the proposed requirement for immediate suspension capability. Both remain proposals; submissions were lodged in October 2025 and the outcome is not yet known.

The definitions side by side

Table 1. How each term is drawn, and what it catches
TermWhereTestStatus
Algorithmic tradingUK — MiFID-derived, MAR 7AComputer determines order parameters with limited or no human interventionIn force
High-frequency algorithmic trading techniqueUK — same frameworkLatency-minimising infrastructure, no human intervention per order, high intraday message ratesIn force
Trading AlgorithmAustralia — CP 386 draftComputer algorithm automatically determines one or more order parameters with limited or no human interventionProposed
Trading SystemsAustralia — CP 386 draftAny system for submitting trading messages into a trading platformProposed
Designated trading representative / automated order processingAustralia — current securities rulesDistinguishes human-keyed from machine-generated ordersIn force; proposed for replacement
No equivalent defined termNew ZealandData not found
AI system, by risk classEU — Regulation (EU) 2024/1689What kind of system it is, not what it does to marketsIn force since 1 Aug 2024

Why "AI" appears in none of them

This is a design choice with a clear rationale, and it is the single most useful thing to understand about the whole regulatory picture.

A rule that named a technology would date immediately. A definition covering "neural networks" would not have covered gradient boosting, and neither would cover whatever replaces transformers. It would also create arbitrage: a firm could restructure a system to fall outside the named technique while doing exactly the same thing to the market.

A functional definition avoids both. Asking whether a computer determines order parameters without a human captures the risk regulators care about — orders reaching the market faster than a person can supervise them — regardless of what produced the decision. The same test caught program trading in the 1980s and catches a reinforcement-learning agent today.

The consequence for the FCA's stated position, on its AI approach page updated 13 February 2026, that it does not plan extra AI regulation, is that this is not regulatory reluctance. The existing definitions already reach AI systems. Adding an AI-specific rule would create a second, overlapping and less durable test.

Two ways to write the rule, and why only one survives Define by technology "Systems using neural networks…" Fails because: — the next architecture is not covered — restructuring escapes the rule — identical market risk, different label — rule needs rewriting every few years Define by function "A computer determines order parameters with limited or no human intervention" Works because: — catches the risk, not the method — covered program trading in the 1980s — covers reinforcement learning today — no rewrite needed when technique changes This is why the FCA can say it plans no extra AI regulation without leaving AI systems unregulated.
The functional approach is the reason the UK and Australian formulations look almost identical despite being drafted years apart.

The EU AI Act sits on a different axis

Regulation (EU) 2024/1689 entered into force on 1 August 2024 and does define AI systems, classifying them by risk. But it is not a trading rule and does not define AI trading. It asks what kind of system something is; the financial rules ask what the system does to a market.

A system can therefore sit in both frameworks at once, in neither, or in one only, and the answers are independent. The UK has not adopted an equivalent horizontal AI statute for financial services, so for UK firms the financial axis is the operative one.

Does your setup count?

Two questions, in order. They are separate and people routinely collapse them.

  1. Does the system meet the functional test? Does a computer determine order parameters with limited or no human intervention?
  2. Are you a person the rules apply to? The algorithmic trading regime binds authorised firms engaged in algorithmic trading, direct electronic access providers and general clearing members.

An individual trading their own money answers yes to the first and no to the second. The system is an algorithm in the ordinary sense and the regime does not apply, because trading your own account is not a regulated activity in any of the three markets we cover.

Table 2. Common setups against both questions
SetupMeets the functional test?Regime applies?
Individual runs a bot on their own accountYesNo — not a regulated activity
Individual uses AI signals, places every order manuallyNo — human determines parametersNo
Authorised firm runs an execution algorithm on a UK venueYesYes — MAR 7A and RTS 6
Authorised firm uses a purchased third-party algorithmYesYes — responsibility sits with the firm, not the vendor
Firm's system only routes orders a human specifiedNoNot on this basis
Robo-advisor allocating client portfoliosDepends on the order mechanicsRegulated as advice or managed service regardless
Vendor selling trading softwareNot applicable to the vendorSelling software is generally outside the perimeter

The fourth row is where the FCA's 2025 multi-firm review found the real problem. Buying an algorithm does not transfer responsibility for it: across the ten firms reviewed, pre-trade controls were adequate, but some users of third-party algorithms did not have a good technical understanding of how those algorithms were built.

Why the everyday words do not map

Four terms circulate in retail discussion, and none corresponds cleanly to a regulatory category.

  • "Bot" usually means any automated trading software, including tools that only produce signals — broader than the regulatory definition in one direction, narrower in another.
  • "AI trading" is a marketing term with no regulatory content whatsoever.
  • "Algo trading" in retail use often means any rules-based system, including manual rules a person executes.
  • "HFT" is used loosely for anything fast, whereas the regulatory term requires three specific characteristics together.

The mismatch matters when reading vendor claims. A product described as "fully compliant algorithmic trading software" is claiming something the regulatory framework does not offer: compliance attaches to firms and their controls, not to software.

What we could not establish

  • Any New Zealand defined term for algorithmic trading. Market participants operate under the NZX Participant Rules, edition 19 February 2026, and licensing runs by activity rather than technology. Data not found.
  • Published supervisory guidance on where the "limited human intervention" line falls for systems with a human approval step. Data not found.
  • The final form of the CP 386 definitions. Submissions were lodged in October 2025; the outcome is not yet published. Data not found.

Key takeaways

  • One test governs everything: does a computer determine order parameters with limited or no human intervention.
  • No definition mentions AI, and that is deliberate rather than an oversight.
  • HFT is a narrow box inside algorithmic trading, requiring three characteristics together.
  • Australia is proposing the UK's wording, which would harmonise two rulebooks that currently diverge.
  • Meeting the definition is not the same as being regulated. Ask both questions separately.
  • Software cannot be compliant. Firms and their controls are; a product is not.

Frequently asked questions

Is my trading bot an algorithm in the regulatory sense?

Probably yes as a description, but that alone does not make you regulated. The definitions apply to authorised firms and market participants. An individual running a bot on their own account is not caught by the algorithmic trading regime regardless of how sophisticated the system is.

How does UK law define algorithmic trading?

Under the MiFID-derived framework carried into MAR 7A, algorithmic trading means trading where a computer algorithm automatically determines individual parameters of orders — such as whether to initiate the order, its timing, price or quantity — with limited or no human intervention.

What is ASIC's proposed Trading Algorithm definition?

CP 386, published 27 August 2025, proposes defining a Trading Algorithm as a computer algorithm that automatically determines, with limited or no human intervention, one or more parameters of an order. It closely mirrors the UK formulation. It is a consultation proposal, not a rule in force.

Does the definition mention artificial intelligence?

No. None of the trading definitions in the UK or Australia refers to AI, machine learning or neural networks. They turn on whether a computer determines order parameters with limited human intervention, which captures a machine-learning model and a simple rule identically.

What is the difference between order routing and algorithmic trading?

Systems that only route or confirm orders without determining their parameters generally fall outside the definition of algorithmic trading. The distinction is whether the system decides the substance of the order or merely handles its transmission.

What does ASIC's proposed Trading Systems concept cover?

CP 386 proposes replacing the existing designated trading representative and automated order processing distinction with a single broad concept of Trading Systems, described as any system for submitting trading messages into a trading platform. It is deliberately wider than the Trading Algorithm term.

Is high-frequency trading defined separately?

Yes. In the MiFID-derived framework, high-frequency algorithmic trading technique is a narrower subset of algorithmic trading, identified by infrastructure that minimises latency, system-determined order initiation without human intervention, and high intraday message rates.

Does New Zealand define algorithmic trading?

We located no New Zealand definition equivalent to the UK or proposed Australian ones. Market participants operate under the NZX Participant Rules, and financial advice and discretionary management are licensed by activity rather than by the technology used. Data not found for a defined term.

Does the EU AI Act define AI trading?

No. Regulation (EU) 2024/1689 classifies AI systems by risk for the purposes of AI regulation generally, not by trading function. It operates on a different axis from the financial rules: one asks what the system does to markets, the other what kind of system it is.

Why do regulators avoid defining AI in trading rules?

Because a definition naming a technology dates quickly and creates arbitrage. A rule about neural networks would not cover the next architecture. Defining by function — does a computer determine order parameters without a human — survives changes in technique.

If I use an AI tool for signals but place orders myself, is that algorithmic trading?

On the definitions as drafted, generally not, because a human is determining the order parameters. The line is human intervention in the order decision, not whether a model informed it. Firms should still assess their own position against the rules that apply to them.

Where can I find these definitions?

The UK definition sits in the MiFID-derived framework carried into the FCA Handbook at MAR 7A, with the technical standard in onshored RTS 6. The Australian proposal is in ASIC CP 386 of 27 August 2025 and its draft rule attachments.

About this page

Compiled by AI Trading Book Editorial from the FCA Handbook, UK-onshored MiFID II technical standards and ASIC CP 386 with its draft rule attachments. Definitions are paraphrased for readability; firms should work from the rule text rather than from this summary. Australian terms are identified as proposals throughout. Items we could not trace are marked "data not found". Published 27 August 2026; last updated 28 August 2026. Corrections are logged on the corrections page.

Sources

  • FCA Handbook, MAR 7A — algorithmic trading; scope covering algorithmic trading, direct electronic access and general clearing members.
  • UK-onshored MiFID II framework — definitions of algorithmic trading and high-frequency algorithmic trading technique; RTS 6 technical requirements.
  • ASIC — CP 386, "Proposed amendments to the ASIC market integrity rules: Trading systems and automated trading", 27 August 2025, with draft rule attachments — proposed Trading Algorithm and Trading Systems definitions; replacement of the designated trading representative and automated order processing distinction; IOSCO alignment objective.
  • FCA — AI approach page, updated 13 February 2026 — no extra AI regulation planned.
  • FCA — multi-firm review of algorithmic trading controls, 21 August 2025 — adequate pre-trade controls; limited understanding of third-party algorithms among some users.
  • NZX — Participant Rules, edition 19 February 2026 — cited for the absence of an equivalent defined term.
  • Regulation (EU) 2024/1689 — in force 1 August 2024, cited for the contrast between risk-class and functional definitions.

Informational research only. Nothing on this page is legal, financial, tax or investment advice. Regulatory definitions are paraphrased and simplified; whether a particular system falls inside a definition is a question of fact and law for which qualified legal advice should be taken.